statistical modelling(upscaling) of gross primary...

52
Statistical modelling (upscaling) of Gross Primary Production (GPP) Martin Jung MaxPlanckInstitute of Biogeochemistry, Jena, Germany Department of Biogeochemical Integration mjung@bgcjena.mpg.de

Upload: buidien

Post on 21-Sep-2018

222 views

Category:

Documents


3 download

TRANSCRIPT

Statistical modelling (upscaling) of Gross Primary Production (GPP)

Martin JungMax‐Planck‐Institute of Biogeochemistry, Jena, Germany

Department of Biogeochemical Integrationmjung@bgc‐jena.mpg.de

Motivation

• Substantial uncertainty of GPP simulations of Land Surface / Terrestrial Ecosytem Models

• Independent, complementary, more observation‐based data stream with least assumptions about ecosystem functioning

• Improved diagnostics, cross‐consistency checks, LSM evaluations

Huntzinger et al 2012

Fluxnet‐Canada

Ameriflux

LBA

CarboAfricaAfriflux

Carboeurope/NECC TCOS

AsiafluxKoFlux

Ozflux

ChinafluxUSCCC

FLUXNET: a network of network of eddy covariance sites

FLUXNET gridded products rationale:

(Direct) observations

(Free-running) C4-models

OfflineDGVM

Remote sensingmodels (CASA,

MOD17)

Empirical ‘models’

Assumptions about systemAssumptions about system

Observational inputObservational input

… be as much as possible on observational side!

remote sensingof CO2

Tem

pora

l sca

le

Spatial scale [km]

hour

day

week

month

year

decade

century

local 0.1 1 10 100 1000 10 000 globalCountries EUplot/site

talltowerobser-

vatories

Gridded Forest/soil inventories

Eddycovariance

towers

Landsurface remote sensing

From point to globe via integration with remote sensing

Tree rings

fSite-level explanatory variables• Meteorology• Vegetation type• Remote sensing

indices

Target variableecosystem-atmosphere flux

The same gridded explanatory variables

Gridded targetvariable

Application

General Principle

‘Proof of concept’

FLUXNET representativeness

Upscaling strategy using a model tree ensemble (MTE) 

Jung et al. (2009) BG

A Model TREE

Advantages:

• ‚Functional‘ stratification

• Intuitive• Can cope with

switches• Seemless integration

of categorical variables

Mimicking a process model (LPJmL)

FLUXNET database

Quality control

Extracting simulated GPP & model forcing data at locations & 

times

Train Model Tree Ensemble

Apply MTE global

Compare to LPJmL simulation 

25 trees, R2 = 0.966

3550 site‐months from 178 sites

R2 = 0.92

Jung et al. (2009) BG

Site level

Global scale

=‚observed‘ =emprically upscaled

Jung et al. (2009) BG

Mimicking a process model (LPJmL)

The value of ‘ensemble magic’

Moving to upscaling with real world observations

Comparison of two flux partitioning methods

Lasslop et al 2010

0 40 80 120160200240280320360400GPP_ust_range_full [gC m-2 year-1]

0.00.1

0.2

0.3

0.40.5

77

118

6148

28 2413 9 8

2 5 1 1 3 2 2 0 3 2 0

Uncertainty due to u* filtering

Reichstein et al in prep

Quality control

• Remove data points with more than 20% of gapfilling

• Remove outliers from the comparison of Reichstein vs Lasslop flux partitioning methods (monthly data)

• Remove 5% of data that show largest uncertainty due to u* filtering (monthly data)

Inputs

Cross‐validation resultsdata from EACH SITE were partitioned into 5 parts 

Site 1Site 2Site 3Site 4Site 5

time… … … … … …

data from entire sites were left out

Site 1Site 2Site 3Site 4Site 5

time… … … … … …

Possible reasons why anomalies are (apparently) predicted poorly

• low signal to noise ratio of monthly anomalies of eddy covariance data

• Large noise in site‐level remote sensing data• Substantial ‘non‐climatically’ caused anomalies in FLUXNET data (e.g. management, disturbances, manipulations)

• Anomalies due to ecophysiological variations that change the sensitivity to meteo and biophysical properties

Global Results

Patterns of Seasonality

Amplitude of mean seasonal cycle

Max of mean seasonal cycle

Hot spots of interannual variability

Towards improving …

The 4 principal dimensions

Injected Information

Data quality Data quantity

Regression Algorithm

Where is the bottleneck?

Sandbox analysis I: Do we use a suitable algorithm?

Sandbox analysis II: Are we limited by the number of sites?

A monte‐carlo test with Random Forests

Aerial photo

Remote sensing data quality issues

Landsat

MODIS

1 km

Towards a coordinated set of experiments and intercomparisons:

Concrete FLUXCOM goals

• To deliver a best estimate ensemble product of carbon  and energy fluxes from an ensemble of diverse data‐oriented and FLUXNET based approaches.

• To study and if possible quantify sources of uncertainty which hopefully leads to improved strategies with reduced uncertainty along the way of the FLUXCOM activity.

• To come up with a practice guidance regarding regionalization of FLUXNET data.

Data base of explanatory variables

P1 P3P2 P4

M1 M2 M1 M2 M1 M2 M1 M2

Consistent cross‐validation & model selection

Analysis and intercomparison of global results

Participants

Approaches

The FLUXCOMP vision

Current Participants

Name Method/Model Approach

Kazuhito Ichii SVM Machine Learning

Dario Papale ANN Machine Learning

Antje Moffat ANN Machine Learning

Gustavo Camps‐Valls ANN Machine Learning

Christopher Schwalm ANN Machine Learning

Martin Jung MTE, RF Machine Learning

Enrico Tomelleri MOD17+ Model‐Data‐Fusion

Nuno Carvalhais CASA Model‐Data‐Fusion

Timothy Hilton VPRM Model‐Data‐Fusion

Anthony Bloom DALEC Model‐Data‐Fusion

Some quick pre‐FLUXCOM comparison (fully unharmonized)

• SVM (Kazuhito Ichii)• ANN ensembles (Dario Papale)• MTE (Martin Jung)• MOD17+ (Enrico Tomelleri)

GPP Mean annual fluxesgC

m-2

y-1

Mean Seasonal cyclesgC

m-2

d-1

FLUXCOM products will capitalize on:

• An ensemble (of ensembles ) of estimates• Higher temporal and spatial resolution• More injected information (variables)• ~twice the amount of La Thuille FLUXNET data

The 4 principal dimensions

Injected Information

Data quality Data quantity

Regression Algorithm

Which and how many variables are needed?

• Injected Information = f(availability, imagination)• potentially large set  feature selection problem

• Consider some trade‐offs … Climate variables

Precise in‐situ measurements

Need coarse resolution climate data for global upscaling

Uncertainty and biases of global climate fields

Satellite variables

No product switch necessary from site to globe

High res global output possible

Very noisy at site‐levelTower satellite footprint mismatch

gaps

# candidate variables

# selected

CLIM+SAT 166 10

CLIM+SAT (ANO) 166 16

SAT 165 9

CLIM 61 7

Some experiments with a feature selection algorithm and Random 

Forests?

Results based on 10‐fold cross‐validation and bootsrapping

On the evaluation of the productsSite‐level cross‐validation is important and valuable to compare different approaches and set‐ups but…

Is a not sufficient evaluation of global fields because:

• Biased sampling of biosphere by tower sites• Noise and site‐specific peculiarities at site‐level• Doesn’t capture uncertainties of global driversNeed for an independent, observational based approach

Fluorescence

If photosynthesis is so complicated why can we get the big picture of gpp relatively easily from sparse data?

- Adaptation to the environment

- At ecosystem scale everything scales with LAI/FPAR

What is the added value of fluoresence?

Does it track physiological regulations that control interannual variability?

Thank you very much for your attention!

Interannual variabilitygC

m-2

d-1

Conclusions

• Global upscaling from FLUXNET is feasible• A new observation driven data stream• Known issues:

– Spatial distribution of mean NEE– Underestimation of interannual variability

• Largest potential for future improvements: adding informative variables

• Plan of a coordinated intercomparison(FLUXCOMP)

• Global products are available on request

GPP

Uncertainties of empirical upscaling

Empirical model/Upscaling tool

Global gridded datasets

In‐situ data (FLUXNET)

Quality control:‐ Gap filling‐ u* uncertainty‐ Consistency of flux partitionings‐ Consistency of energy budget

‐ Avoiding problematic variables (e.g. VPD)‐ Identify & use good quality products‐ sensitivity studies with various forcings

‐ use various approaches‐ ensemble methods‐ artificial experiments

GPP anomaly snapshots (gC m-2 mo-1)

Jung et al. in prep 52